YOLO vs. diffusion networks for underground pipe detection: a case study using ground penetrating radar data

Abstract The accurate detection and mapping of underground pipelines are critical for urban planning and infrastructure resilience. This process is critical to prevent catastrophic failures, minimize service disruptions, and maintain the structural integrity of public works projects. Ground Penetrating Radar (GPR) is a prominent technology for the automated detection and precise localization of subsurface utilities. However, the current dependence on manual data interpretation is a major bottleneck, often resulting in high temporal and operational costs. Consequently, this research addresses these limitations by presenting a performance analysis of two distinct deep learning architectures–YOLO (You Only Look Once) and Diffusion-based neural networks–for automated underground pipe detection. Utilizing a specific dataset of GPR scans, the models were evaluated using a robust five-fold cross-validation framework to assess mean Average Precision at 50% Intersection over Union (mAP50) and inference latency. The experimental results demonstrate that YOLO-based architectures provide a superior balance between detection precision and real-time processing efficiency. While five-fold cross-validation established a mean mAP50 of approximately 80% for the YOLO models, the YOLOv12n architecture exhibited superior performance on the independent test set, achieving an mAP50 of 94%. In contrast, DiffusionDet models were characterized by lower precision and significantly higher latency. While this study focuses on a single-site case study, the findings establish a foundational baseline for the selection of deep learning architectures and provide a framework for future research into the adaptation of these models across varying geological conditions.